This is Part 2 of The Real Cost of Agentic AI series.
Summary
Microsoft’s Copilot introduces “Agent flows” – a new type of workflow running within AI agents – which exist alongside traditional Power Automate “Cloud flows”. This post explains the differences between agent flows and cloud flows, how Microsoft has put them under separate licensing buckets, and why this split adds yet more complexity to the Power Platform licensing story.
What’s happening
With the advent of Copilot and generative AI, Microsoft is reimagining automation by letting AI agents orchestrate actions. Enter Agent Flows – lightweight automations built in Copilot Studio that a Copilot (AI agent) can execute as part of a conversation. For example, if a user tells a Copilot agent “Update the CRM record for this customer,” an agent flow could be triggered behind the scenes to perform that update. These agent flows are created using natural language prompts or low-code interfaces within the Copilot Studio environment, embedded in the AI’s dialog flow rather than a standalone workflow.
By contrast, Cloud Flows (the classic Power Automate flows) are the established automation processes triggered by events, schedules, or APIs, used for enterprise-wide tasks like approvals, data syncs, and integrations. Technically, both are automation pipelines – in fact, an existing Power Automate flow can even be converted into an agent flow in some cases, bridging the two worlds.
However, Microsoft has drawn a clear line in terms of licensing and usage. Agent flows fall under Copilot Studio’s licensing model, meaning they consume Copilot capacity (measured in “messages” or actions tied to the AI agent). In other words, if your Copilot agent runs an agent flow, it counts against your Copilot’s metered usage (or message pack). Cloud flows fall under Power Automate licensing, requiring either per-user licenses or a special per-flow license for runs by unlimited number of users.
The key difference is that agent flows are considered part of the AI experience you pay for via Copilot, whereas cloud flows are part of the Power Platform you license separately. Microsoft has essentially duplicated automation functionality in a new context – and with it comes a duplicate licensing schema.
Additionally, there are practical distinctions: agent flows are optimized for chat-based, single-user scenarios (not easily shareable or reusable across the whole org), whereas cloud flows are enterprise-grade workflows that can be shared, versioned, and managed centrally. In summary, we now have two parallel automation tracks inside Microsoft’s ecosystem – one native to AI agents, one in the traditional workflow engine – each with its own cost structure and ideal use cases.
Why it matters
For anyone working with Microsoft’s business apps, this dual model raises important questions and potential confusion. If you’re a solutions architect or a power user, you might be wondering: When should I use an agent flow vs. a cloud flow? And am I paying twice if I use both? These questions matter because choosing the wrong approach could mean unnecessary costs or limitations.
For example, if you have an existing Power Automate cloud flow and you attempt to trigger it via a Copilot agent, you’ll need to have both a Power Automate license (for the flow) and Copilot capacity (for the agent calling it). You don’t want to be double-billed for automation just because it’s AI-assisted. Conversely, if you rebuild that process as an agent flow, you might save on Power Automate license costs, but then those runs will draw down your Copilot message quota – which might become more expensive if the flow is complex or runs frequently.
This is a classic Microsoft licensing puzzle: two paths to automation, with different meters running in parallel. It also matters for governance and maintenance. Cloud flows are well-established in many organizations with proper DevOps, ALM (application lifecycle management), and sharing capabilities; agent flows are new and cannot be shared or reused as broadly (they’re tied to the specific Copilot agent’s context). That means if a business process grows beyond the AI chat scenario, you might have to refactor an agent flow back into a standard flow to scale it – essentially re-doing work.
From a cost perspective, companies need to evaluate the trade-offs: agent flows might be cost-effective for embedding simple tasks in AI conversations, whereas cloud flows might be better for heavy-duty, cross-team automations (where you’ve perhaps already paid for unlimited runs via a per-flow plan). All of this adds another layer of complexity to an already complex Power Platform licensing landscape.
For IT managers and partners, it’s crucial to prevent confusion: users shouldn’t create agent flows willy-nilly thinking “it’s free with Copilot” when in reality it’s burning through Copilot credits. Likewise, you don’t want to restrict innovation by banning agent flows entirely – they have their place.
The bottom line: it matters because it forces organizations to consciously choose how they automate, with licensing considerations now in the mix of that architectural decision. (I often help clients draw these lines – figuring out which automation should live where – to avoid unpleasant billing surprises and ensure maintainability.)
My perspective
I have mixed feelings about this bifurcation of flows. On one hand, Agent Flows are a smart innovation – they make AI agents truly useful by letting them not just chat, but act. That’s a big deal for productivity, and in my personal hands-on trials, building an agent flow with natural language is almost sci-fi cool. Microsoft is rightly proud of making automation more accessible (who wouldn’t want their Copilot to do things, not just talk?).
But as an advisor who has witnessed years of Power Platform licensing confusion, I can’t help seeing a familiar pattern. Microsoft has essentially created yet another silo of functionality with its own rules. It reminds me of when they introduced Power Apps Portals or AI Builder – new capabilities that at first seemed separate, then had to be reconciled with the existing models. Here we go again: “Copilot Studio included, Power Automate extra”, “message-based vs action-based limits” – try explaining that to a project team without their eyes glazing over.
My somewhat contrarian view is that the true difference between “process automation” and “AI agents” might simply be the licensing meter attached. Microsoft is effectively nudging us toward the Copilot/agent side by bundling some automation there. Today they say Power Automate “isn’t going anywhere”
– and I believe that, for now. But if agent flows take off, I wouldn’t be surprised if more and more workflow scenarios (especially those initiated by conversational interfaces) get funneled into the Copilot Studio universe. That could mean eventually a consolidation – or a deprecation – who knows?
In the meantime, my take is pragmatic: use the right tool for the job. If an automation needs broad reuse, robust logic, and you’ve already paid for a Power Automate license, stick with a cloud flow (Copilot can perhaps invoke it indirectly via an API call, staying within one licensing realm). If it’s a contextual, one-off action in a chat, agent flows shine – just keep an eye on how often they run. (In planning sessions, I often recommend a cost-benefit analysis of agent vs. cloud flow for each use case – one of those things an experienced licensing advisor can assist with.)
Ultimately, Microsoft’s dual approach gives us flexibility, but also enough rope to hang ourselves if we’re not careful. Let’s enjoy the new tech, but mind the licensing fine print while we do.
